Malevolent Cartography

Write C

Drink your coffee black

Sleep on the floor

Burn your rolling luggage




Setjmp/Longjmp Exception Handling in C

2020-03-25 by Quinlan Pfiffersoftwarewritingc

With quarantine being a Real Thing in our lives now, I’ve been spending a lot more time on personal projects than I have in a while. One of the things I recently added to Lair is a sort of runtime exception. I needed something like this because without it, I wouldn’t be able to properly do unit testing on real code samples. I was just calling exit(1) before, which is not really useful.

I’m always on the lookout for fun and weird new syscalls, and setjmp(3) has been in the back of my mind for a while. Reading the wikipedia page can be tremendously useful here.

The way I like to think about setjmp and longjmp is that they’re like contextualized gotos, with similarities to fork(). That is: you call setjmp() and evaluate it’s return value in a branch, just like you would with fork, but you can actually re-enter that point from pretty much anywhere you have access to the jmp_buf you defined when initially calling setjmp(). Let’s look at how it’s used in Lair.

Lair has a very small struct _lair_runtime object associated with each running program that currently only stores exception data. It’s definition is something like this:

struct _lair_runtime {
    ERROR_TYPE exception_type;
    char *exception_msg;
    jmp_buf exception_buffer;

It’ll probably change in the future, but for now it’s sort of like a poor-man’s perror(): You have one exception, and it’s information/message is stored in that runtime object when it happens. The real interesting part to the runtime data is the jmp_buf exception_buffer; though: That’s where the contextual data for the exception is stored. Here’s how it gets setup in the lair_execute() function:

int lair_execute(const char *program, const size_t len) {
    struct _lair_runtime *runtime = _lair_runtime_start();
    if (setjmp(runtime->exception_buffer)) {
        if (runtime->exception_msg) {
            print_error(runtime->exception_type, runtime->exception_msg);
        goto error;
    /* ... */

This is that conditional evaluation I was talking about before: the first time you evaluate setjmp(), it’ll return 0 but store it’s current context into the specified jmp_buf. This let’s you do cool things later on, which we’ll talk about shortly. Since it returns 0 initially, we move on and execute whatever program we’re trying to run.

Later on, when we encounter some unforgiveable condition like a syntax error, parsing error, undefined function, etc. we can call this code:

void throw_exception(
        struct _lair_runtime *r,
        const ERROR_TYPE err_type,
        const char *msg) {
    r->exception_type = err_type;
    r->exception_msg = strdup(msg);

    longjmp(r->exception_buffer, 1);

There is also a handy helper function for checking runtime constraints called check() in Lair that will wrap this in a simpler way, so we can do something like this (as an example):

check(r, ast->indent_level > initial_indent_level, ERR_SYNTAX, "No 'True' condition to follow.");

Anyway, the important thing to notice here is the longjmp() call. longjmp() takes a jmp_buf and a value as an argument. The jmp_buf is important because it tells longjmp() where to go, and the value is important because it tells setjmp() what to return: if you hand longjmp() a 1, setjmp() will return 1 waaaaay back in our initial conditional, which for review is here:

    /* ... */
    if (setjmp(runtime->exception_buffer)) {
        if (runtime->exception_msg) {
            print_error(runtime->exception_type, runtime->exception_msg);
        goto error;
    /* ... */

The only caveat to this is that if you pass longjmp() a 0, setjmp() will behave as if it returned 1. Basically setjmp() will resolve truthily regardless of what number you pass to longjmp(): You just get to pick what value of truthiness.

So now that we’ve called longjmp(), our instruction pointer is now back at this setjmp() call, except it acts as though it’s returning whatever we handed to longjmp(), which in our case is 1. This means we enter the branch, check our exception_msg, print it out and die! Nice. This is much better than just calling exit(1), now we can run multiple unit tests as black boxes and expect their results.

This isn’t “real” exception handling yet, but is an excellent starting block. Things to consider when moving forward are:

  • Storing your current stack frame in your runtime object so you can bubble back up through the call stack
  • Catching exceptions and handling them in your language
  • Exception types
  • Exception classes and catching them: eg. Should you be able to catch a runtime parsing error?